The Radeon RX 6900XT Buildzoid Edition.

hi guys bill zoid here and today we're going to be taking a look at my rx 6900xt red devil ultimate buildsoid edition which um is probably the fastest air cooled 6900xt in the world um so let me show you some scores to sort of back up that claim because that is a big claim to make um so first up we have uh my fire strike ultra score um right here and yeah that's that's on air cooling and um yep i'm in fourth place with a 6900 xt um and the fastest 6900xt currently on hardware ball i barely beat the barely beat the guy under me uh but he's on watercooling so it's just like like yeah but like anyway like i beat him right and i'm on air cooling so because the thing is like with water cooling it's like i technically could water cool the card um i do have the ability to set that up the thing is um like actually like looking at his score it doesn't look like it would make much of a difference to my operating temperatures because he's hitting 49 degrees celsius and i was hitting like 58 so um yeah that's like not worth the effort as far as i'm concerned and also if i'm gonna go through all that work then i may as well go all the way with liquid nitrogen um and the thing about the whole well bills lloyd why haven't you run the card on liquid nitrogen yet well you see i bought a 6900 xt i contrary to popular belief power color didn't send this thing to me um which actually i'm not sure that anybody thinks power color sent me the card but anyway if you thought they did they didn't i bought it i'm very upset about that because this thing has garbage input filtering um like yeah anyway so so i bought it and i plan to spend more money on like getting it into a better condition than it comes in from the factory which is why it's already like it's already in a better condition than it comes in from the factory um from a overclocking perspective not necessarily a reliability perspective i don't really trust my soldering skills to not like have well i've not yet had a capacitor fail on me due to how i've soldered it but from reading capacitor documentation my soldering technique is definitely not what would be considered like safe for long-term use so from a reliability perspective i think we can say this is probably worse but from an overclocking perspective it's like it's clocking better than when i first got it even at the exact same voltage but anyway so yeah fire strike ultra um you know top score on well yeah top score for a 6900 xt and i'm i'm just on air cooling and i do plan to take the card cold eventually it's just that my budget needs to recover from the um well damage that it has uh suffered from the purchase of a 6900 xt anyway um then fire strike extreme i'm in second place for fire strike extreme um so that's an even better result now funnily enough you can't run like i can't run fire strike extreme at the same settings that i can run fire strike ultra like ultra runs at a higher frequency or at least um the the awkward thing is like i'm actually not sure what the readout is from from 3dmark but in terms of the driver set like the settings in the amd overclocking menu um like i'm running this at higher settings than i am running this now the reason i say higher settings and not higher clocks is because like this does not reflect my experience of how stable it was at all like there is no way it was running hundred megahertz lower cause on average because fire strike extreme was very very like it would just crash if i was running the same settings as i was running for ultra so i think the clock readout on on navi is just completely misleading um and should just be ignored like i wasn't looking at the clocks when i was benching it i was literally just looking at the scores um so yeah anyway fire strike extreme is actually harder to run at least for me i don't know how it is for other people but for me definitely fire strike extreme was sort of like wait a minute this this can't run the same clocks that i can run for ultra my theory on why that happens is because the ultra runs at lower fps and so the transient loading that you get from the power draw fling like because each time the card goes through a frame render cycle different parts of the frame render cycle have different levels of power draw so you get like the the transient loading of like fire strike extreme happens 200 times a second right because fire strike extreme runs at around 200 fps whereas fire strike ultra well fire strike ultra runs at around 100 fps much slower than fire strike xtreme does um and actually in some parts it drip dips all the way down to like 50 fps fire strike fire strike ultra is fire strike at 4k so because it runs at much lower fps you're going through that transient loading 50 times a second instead of 200 times a second and my theory is is the inadequate input filtering capacitor configuration on this card has much more time to recover on a 50 time like like can deal with something that happens 50 times a second a lot more easily then it can deal with something that happens 200 times a second so yeah i'm i'm currently in the in the process of searching for extra capacitors to to put on the card which is uh why i have two digikey tabs open but um anyway so yeah fire strike extreme second place so that's even like a better ranked result because the 30 90 is less competitive at lower resolutions com to the 6900 xt and the 6900 xd sort of loses performance as you get higher and higher resolution because the infinity cache gets less and less efficient as the resolution gets bigger um but anyway uh and then i also ran fire strike extreme for hardware bot now the funny thing is i actually ran this this score like i got this score before i went and also did ultra and uh extreme for the 3d mark hall of fame you might be like why why are you why do you bench these separately well the thing is hardware bot has a different rule rule set from the 3d mark hall of fame so for 3dmark hall of fame uh you can't have like a warning like this because this this is no tessellation but on a hardware bot that's legal just because of how the hardware bot database works basically for a long time 3dmark couldn't detect if people were disabling tessellation um but eventually the people who run 3dmark decided hey we're like added detection for tessellation and lod tweaks and all of that so now if you run 3dmark you have to run it with tessellation at stock like 3dmark settings you can't change the tessellation settings and you also can't modify the lod or anything like that you can't do any of these these tweaks that you could do for hardware ball now and when 3dmark made this change over to like detecting tessellation they deleted all of the old scores like yeah i think they deleted basically all of the old scores because it was just like well we can't prove that these weren't done with tessellation disabled um or maybe they did delete them like i'm not sure how it was handled but basically uh they chose to purge uh scores that would be without tessellation um the hardware bot team however sort of decided like the values the integrity of the database more so than the integrity of the scores themselves so on hardware bot we still get to bench no tessellation because all of the fire strike submissions for like several years didn't run tessellation all of the top scores were without tessellation so instead of removing all the old top scores it was just kind of like we'll just keep running with no tessellation so now you bench hardware bot you disable tessellation if you bench 3dmark you turn tessellation back on um doesn't really like i i really don't have a problem with it right like if everybody's using the same rules you know it's fine um so anyway um but yeah so i did did this this run first and this was actually at lower settings because i normally prioritize hardware bot in terms of benchmarking like i don't really pay attention to the 3dmark hall of fame much at all um but then when i was sort of thinking about you know what i want to do for this video i was like you know what maybe i should run the card with tessellation turned on so people don't get confused by like why is my fire strike right like score so big it's like yeah it's because there's no tessellation in that um but uh anyway and this was run at lower clocks because when i ended up running fire strike ultra like this is way more gpu bound than uh like with with a score like this the cpu plays a bigger role than a score for fire strike ultra so for fire strike ultra i really needed to max the gpu um in order to like you know because the cpu wouldn't carry it as hard um not that you know like like the next what like four scores below me are all actually like yeah the next several scores are all on 5950xs so it's not like i really have an unfair advantage but um yeah here you really need like for ultra you really need to push the gpu whereas in extreme um you know the cpu can carry you a bit harder because the the score is overall like the the cpu component of the score is bigger than in in ultra um but anyway hardware bot i got fourth place and the three scores above me are all on like sub-zero cooling um actually i think even one of the well yeah sub-zero cooling because the score below me is like sub-ambient because the gypsy over here is using a chiller um but yeah and i'm on so chiller water cooling though you can get chillers that go below zero admittedly once you get those kinds of chillers you're running more like antifreeze than water but still you know technical technically like yeah like you can technically get those sub-zero but anyway so we have an ln2 score above me then a phase change score and then another ln2 score above me so yeah for an air-cooled card probably the fastest one in the world anyway now let's take it apart and talk about what i've done to it and actually before we even start taking it apart the most useful modification you can do to a 6900 xt um is add an i squared like add a nice like a header to hook up a elmore evc2 um which the evc2 is no longer um no longer produced but there's like a successor model to it which has like the same functionality and i think it's slightly cheaper slightly smaller something like that anyway what this what this is is um it's a little controller from elmore um elmore of well from elmore labs and basically you just connect this to a compatible gpu or while more like compatible voltage controller it's not specifically for gpus it'll talk to any voltage controller that this supports and so the 6900xt uses an xdp e132g5 and it also uses a ir35217 both of those are supported by the evc2 so we connect it over uh to here where there's the i square c header for for the card and then you can use this to set the voltages on the xdp e132 g5 and the ir35217 and the reason why this is the most useful modification for competitive overclocking is because this thing even on air cooling it scales um so for the fire strike ultra run i was running plus 88 millivolts v core um so my max voltage was like approaching 1.3 volts at that point um card was also pulling close to like gpu z was reporting like 500 watts of gpu power draw for graphics test one which that's the really really hot one in fire strike so yeah um card scales even if you're on air cooling it scales with voltage so you really want an evc and on all of the cards that use a sort of reference style pcb so like this is a reference style pcb as far as i'm concerned power colors has added some extra phases and they stretch the vrm out a bit but ultimately this is a reference card as far as like this is basically a reference card um and then there's some cards which are literally a reference card with just like padding piece like pcb padding up top and to the side and whatever um but anyway that way like the thing is a lot of the cards that use the sort of like that are based on the reference pcb they're gonna have a a header oh like the pads for an i square c header right here so yeah um you know i just added my own little connector to that and uh now i can attach the evc to it um technically like other people who have run the car like i'm not the first person to hook up the evc to one of these cards um but most of the other people who have been running evcs on these cards they hook up to like the test points coming right off of the voltage controllers so they solder wires directly to the test points and it's just like that's like the uncivilized way to do this okay like the lovely people over at amd uh gave you an undocumented um very actually hard to recognize header right here because it looks like a transistor footprint um but no it's a header it looks like but yeah but admittedly very like annoying to identify uh i square c header like use that for your connection like don't especially like the main reason i don't like soldering wires to test points is that's how you rip traces out of pcbs whereas this is like this is meant to have a connector attached this isn't or wires or anything like that and now admittedly you can glue the wires to the pcb but now you're adding glue to your card like i said uncivilized so yeah that technique works fine this is in my opinion way better and what's kind of neat about the way i actually soldered this this connector you'll notice that it's like really sitting like almost flat against the pcb it's not quite there's a slight angle to it um if it was flat then there's like a little smd component that you you wouldn't get past with the connector um now anyways with how low this is sitting i can actually put the back plate back on the card and that won't touch the back plate like it it clears under the back plate so that's kind of just like like i like to like i like to when i mod cards um like sometimes i don't care but sometimes like i think it's always neat when you can mod a card such a way that it's impossible to tell that it's modified until you take it apart like if i put the back plate back on this you can't really see any of the modifications because they're all under the back plate there's nothing really sticking out anywhere um but yeah that's probably going to change because i can't figure out what to do about the input filtering situation on this card it's it's really quite bad but yeah so attach i square c header over here best place to do it as far as i'm concerned um and yeah so that's super convenient and this is like reference anything reference based so like oc oc formula has it because it's reference based probably the um well i'll have a video about the pcb of the oc formula eventually but yeah so there's the red devil cards that use this layout there's the tai chi from asrock that also has this header there's like a gigabyte card that's also reference based like basically tons of cards have that header right over there so best place to hook up the ice cream c as far as i'm concerned because you really like don't have to worry about it too much if you hook up over here whereas over here it's like very fiddly to do and and annoying and like i'm really surprised nobody bothered to trace this out sooner because it's like there's a limited number of headers on this thing right like we have these um which i checked those none of those knows how i i square i see on them there's this that doesn't do anything there's this that doesn't i don't know what that's for i don't know what this is for i don't know what these are for those aren't i square c headers like this would be like where you'd expect it to be because it's like near the controllers but nah that doesn't i don't know what that's for either it's not for not for the i square c um it's over here but yeah anyway um so that's probably like the most important modification you can do um and then you just buy yourself an elmore labs evc2 and bam you get voltage control technically you can also use a trimmer to skew the voltage sensing circuitry of the voltage controllers i would strongly recommend you don't do that but you can and it can work but it's a terrible idea um like you really don't want to be doing that but it is like like it is possible i don't recommend it um and i'm saying it's and i'm not gonna tell you how to do it either like i've done that but i don't recommend it based on experience so anyway um i'm not sure do we take the card apart now or do we just ah no no no let's not take the card apart just yet well i mean i can start on that um but zoom in a bit so obviously the other thing other than adding the the header over here oh there you can see it a bit better i've gone absolutely to town on this thing adding capacitors just everywhere um so um yeah v core has um so all of these capacitors over here that i added uh those are 22 microfarad um multi-layer ceramics i believe rated for four volts off the top of my head i believe i bought like 200 of them or something their o603 package size they're awesome because o603s like they're very like a small high capacitance convenient to just jam sort of anywhere which is why i've gone ahead and put so many of them on the card um like they just fit they fit really well you don't really have to worry like if there's a little gap you could probably put one um but anyway so i added a bunch of those um which actually didn't really achieve much of anything for for v core at least not initially because the issue is i can't really do like since i have one card i am not gonna go especially since that one card is very expensive i am not interested in like putting stuff on taking it off and then put it back on again for the sake of getting like really good test data because like unless it's getting worse i'm i'm cool with it right like that that's the thing is just like so i don't know if um because the the thing is like initially because i did all of this first because normally my assumption with any gpu is that you can you upgrade the output filtering capacitors first um and then input filtering is like not relevant like the vast majority of gpus you do stuff to the input filtering it doesn't make a difference um this card is the anomaly in my experience where it's like nothing happened until i added a bunch of capacitors to the input and this is still not enough as far as i'm concerned but anyway so yeah all of these 22s for v core i added them to vddci this is vddci over here i added some to 0.75 volts actually i think i added just one to 0.75 volts over here this capacitor didn't used to exist there was an empty pad here so i added that that's for 0.75 volts that was probably a total waste of time that rail i like i don't know that rail is probably very minor but it doesn't hurt anything to add like adding caps doesn't hurt especially when you have empty pads indicating that at some point they were there was meant to be caps there um and then they decided to get rid of them for cost-saving reasons um but anyway then we have the soc rail which i just added some more multi-layer ceramics to that because unfortunately there wasn't any like empty bulk capacitor pads that i could have soldered more capacitors to easily um then v core we have like more multi-layer ceramics on vcor up over here i didn't do anything in this area because i there's not really a lot of space to work in so i didn't didn't really go there um and i don't want to stack them vertically on the back of the card because vertically stacked multi-layer ceramics have very little structural like rigidity so they tend to break off and and cause more like they tend to cause damage like if you put the card down on the desk bad um so yeah so that's sort of what i what i did for the output filtering is just throw 22 microfarad multi-layer ceramics everywhere until i was bored i v core is well in excess of like 30 having 30 of the 22s um the other rails i don't know like literally i just went like i didn't go like hey i'm gonna add x i i just kind of went like until i get bored other than that the memory uh also has some extra capacitors on it and that's like these guys um that's this right here right like you can see that there's there's two of them right next to each other um and this is sort of inspired by nvidia um because if you look at how nvidia does their filtering for their memory chips on their gddr6 cards and then you compare it to how amd does it um i can't believe this actually runs like that's the thing like there's very little capacitance behind the memory chips on a amd gpu compared to an nvidia gpu and so one of my theories for why these cards are so bad at memory overclocking is just like there's not enough filtering for the memory um and so that's why i added all of these uh these these are o402 4.7 microfarad multi-layer ceramics and i added four of them to each each memory chip right so you have one two three four and i did that for every single memory chip um i also added some 22s to the memory rail but those are on the front of the card because there's not really anywhere to put them on the back of the card um and i also added some for 417 microfarad uh polymers to the memory rail as well and i'd say that's responsible for my ability to bench this card at 2180 memory at lower core frequencies i was actually getting like decent like proper memory performance all the way up to like 2 200 megahertz but i assume as the core gets faster the load on the memory controller increases so much that the memory stability gets worse and the thing with the navi cards is that as the memory stability gets worse you initially lose performance you don't initially get artifact and you just start losing performance um which yeah i'm not sure how that works like some people say it's ecc but i don't believe it because if you keep pushing the frequency high enough eventually you start getting artifacting and then like full-on just crashes and just yeah so like these cards generally don't artifact but you can get them to artifact you definitely can so i'm not really like i think um like looking at some of the software monitoring utilities for the the clock speeds on these cards it kind of looks like the cards have dynamic memory clocks as well um i'd love like i really wish i had an amd contact so that i could confirm that but it looks like these have dynamic dynamic memory clocks so maybe when you push the memory clock too high um and it starts getting unstable it starts like throttling the memory clock really hard which would explain why you don't actually see any artifacts but lose a bunch of performance out of just nowhere and it is a lot of performance so like i think the card starts like aggressively underclocking the mem like it might start aggressive like the the clock might hit a like peak frequency get unstable get down clocked like way more than necessary you lose a bunch of performance to that and then it brings back up to peak frequency which is why eventually you can still get it to crash right because if it was if there was errors getting corrected then it'd be like like even at high frequencies you would assume that it wouldn't let errors just pass through into the render output or like completely drop the render output which i had to happen on one okay like before all of the capacitors what was kind of funny is like you could get the card to actually run benchmarks without rendering anything if you just clocked the memory high enough it would still you know the fps wouldn't go through the roof so evidently the core was still processing some kind of data it was just never like passing it to the display output so you never got to see it um so that was that was kind of fun um you can also get that effect on like 290 x's if you push them really really hard and actually a lot of a lot of amd gpus will actually just do this thing where if you keep pushing like if you push them far enough they'll just stop rendering whatever the hell they're supposed to be rendering so you'll just get a black screen and you'll probably still have an fps counter which is really interesting i guess 2d elements somehow don't get affected by it like the 3d part like the 3d part doesn't get handed off to the render out like to the display output but the 2d stuff does so you still get like an fps indicator but it's just a black screen with an fps indicator and you can still and from the fps indicator you can tell like evidently it's not like bugged out and not rendering anything because the fps doesn't shoot through the roof it's just yeah so um anyway um the the memory overclocking situation is interesting but my card like for benching i like all of my top scores are at 21.80 on the memory um which is the highest memory clock i've seen anybody bench one of these with and actually still get good scores like i can go up to 2200 without crashing well sometimes without crashing that's the funny thing is like 2200 sometimes takes out the drivers and sometimes it doesn't um but yeah i can sometimes like if i go up to 2200 it'll definitely perform badly unless i'm at like low core clocks um so yeah like the memory situation on these is is interesting but um yeah so i'd say that the caps are mostly responsible for my ability to bench 2180 i also messed around with like more memory voltage and less memory voltage and more memory controller voltage less memory controller voltage and it doesn't really seem to do anything in fact i ended up running some of my fire strike runs with under vaulted memory just to see if it like does anything to the memory stability it doesn't you can undervolt the memory and it still runs the same clocks and it doesn't like lose performance or anything so super weird same goes for the memory controller like i ended up just running the memory controller at a fixed 0.8 volts instead of leaving it on the dynamic voltage mode where like at idle it would sit at 0.7 and then ramp up to like 1.0.85 i would just run it at 0.8 um but anyway um yeah so the only thing that really seems to like make these cards uh well on air cooling at least the only thing that really seems to be worth doing is just mash the v core um and that that'll make it run higher clocks just fine so that that's how i got the the really high clocks that i'm getting it's entirely just lots and well it's not even a ton of equal it's like point uh like i think i was all the way up to plus 88 millivolts and what's interesting at plus 88 millivolts has finally started to get some artifacting so evidently like well yeah so you can get even the core to artifact which is interesting because before at lower voltages um the core would never artifact it would just crash um so yeah these cards are kind of interesting um to mess with very different from what i'm used to where it's just like okay now it's artifacting but the performance keeps going up now it's already affecting more but the keep the performance keeps going up and now it doesn't run anymore right that's that's the usual expectations for most gpus for me and this yeah this doesn't do that at all this this will quite happily look like it's perfectly stable and then just drop and sometimes it'll like reset the whole system and sometimes it won't which is kind of interesting yeah but anyway the other modification that i did and that made the most difference to the overclocking on this card is not all the output filtering capacitors well i'm not sure maybe like the the thing is um the output filtering capacitors might have been made significantly more effective by the input filtering modifications that i did so you see all of these big capacitors and then also these guys over here which you can also see you know on the other side of the card as well um those are the input filter like the large input filtering multi-layer ceramic capacitors from an rx vega yeah and adding all of those even at stock clocks i mean stock vaults made the card clock like 20 megahertz higher now the problem is i added these after i did all of the output filtering stuff so or at least like 90 percent of the output filtering stuff and then i did the input filtering and then the card started to clock higher and that was and the annoying thing with this is is like so i don't know if like the out like if these played into the effects of the input output filtering or if like if if doing the input filtering upgrades alone might not do anything you might need both input and output filtering upgrades in order to get the effect that i got but yeah like the input filter like because without the input filtering the car just didn't seem to well the memory improved but everything else was just kind of like the core wasn't getting better or anything like that and then i added all of these input filtering caps and suddenly i can run you know i was benching like 28 50 max on stock vaults like what and by stock volts i mean the voltage slider maxed out at 1.2 volts because like it like when you enable the voltage slider or advanced mode or whatever the hell it's called when you enable the voltage slider it defaults to 1.2 volts so i consider that stock okay because it defaults to that but anyway um yeah so i could go all the way whereas before all of the input filtering mods i could only bench like 2820 max clock so i could bench 2820 and then after the input filtering mods which you know after the input and output because the output stuff was already there um i could suddenly bench 2850 um or at least like what 2850 was a bit temperamental it wouldn't always pass but 2844 was like perfectly stable which was like whoa the input filtering on this thing sucks and it really does if you compare the input filtering capacitor configuration of this thing to like older amd gpus like you need to go all the way back to like mid-range cards from 10 years ago in order to find cards with less input filtering than this does than this has and i don't really well no i do fault power color for that and i also fault amd for that because power color took the reference amd design and they stretched it and they didn't do anything to the input filtering they didn't like they didn't improve it in any way shape or form they didn't upgrade it they didn't like downgrade it or upgrade it they just didn't touch it they just kept it exactly the same and then well i guess you could say they added extra multi-layer ceramics but that's because each phase has to have its own uh high frequency you know uh capacitors like right next to it so that they function um so i don't really count these like the multi-layer ceramics that the cards came with for input filtering i don't count those you don't really have a choice with those you you can't not have those as far as i'm concerned um and anyway so yeah you know throwing a bunch of capacitors from a vega on the input filter um made the card clock significantly better but it gets worse believe it or not because after the run in with the sort of like whoa adding a bunch of multi-layer ceramics from a vega makes this card clock uh clock better um i wonder if the bulk capacity like the bulk input filtering capacitance is also that bad and turns out it is um so let's go on to this image right here give me a second there see that so this is a configuration of the card that obviously is no longer like i took that off um which i'm which the problem with that configuration is those caps are gigantic they are 10 millimeters in diameter and i don't want them on the back of my 1800 quid card they're way too big there's also not really anywhere to attach them properly so right there they're just kind of hanging off of the multi-layer ceramics which is just super sketchy if you like if you grab the card wrong or you like forget about it and put something on top of it you're probably gonna break the like tabs off of a multi-layer ceramic with doing that by doing that with caps that big attached to them so that right there is it totally not okay now you could admittedly like glue them to the card which would kind of reduce the risk because again like the the issue is like you'd still have the option like the depend i guess depends how hard the glue is but if the glue is kind of gummy you still have like the torque like you could still torque the the caps um relative to the multi-layer ceramic and it's just like so they wouldn't necessarily come off completely but they would still be able to apply enough leverage to the the caps that they're attached to to cause damage so that definitely just not a viable config like that that is not a configuration of the card that i want to have it totally works um so that further like that didn't have as much of an impact as the the vega caps or at least i didn't test it that much but it slightly helped with core overclocking again but the crazy thing about doing this was that it massively reduced the coil whine which was not something i was expecting whatsoever like that wasn't even what i was testing for i just noticed whoa the card is like way quieter than i'm used to and then it's like oh right um because yeah this thing has a bunch of coil wine and all of the coil wine as far as i can tell is probably from the input filtering inductors and so yeah if you add a bunch of extra capacitance to the input filter on the output side of the input on the like vrm side of the input filtering uh inductors then the voltage swing acro like in front you know like the yeah like the voltage difference across those inductors is going to get smaller therefore the like current flow like the changes in the current flow through those inductors is also going to get smaller therefore the like the the change in the magnetic field of them is going to get smaller therefore they're going to vibrate less therefore they're not going to make as much noise like yeah it makes perfect sense that this would reduce coil wine it's not what i was aiming for but neat it makes less coil wine anyway these caps are 330 30 microfarads each with a seven million esr um and they're way too damn big so i took them back off again but that right there is kind of like so basically after doing that test because the reason i did that test was like i already initially had a suspicion that the input filter like after the multi-layer ceramics from the vega i was like oh okay so the input filtering on this is evidently quite bad i wonder if not only is it lacking in the multi-layers like in in the the high frequency stuff and i don't even know that it was lacking in the high frequency stuff um because yeah i added a bunch of multi-layer ceramics but i also added a bunch of them so there's a decent chance that the improvement was just from the extra capacitance alone rather than the reduction in impedance that a bunch of extra multi-layer ceramics in parallel would offer um but uh yeah so just adding a bunch of capacitance on its own might have been the benefit so uh i wanted to run an experiment would more capacitance make the card better like a lot more capacitance make the card better and yeah it does um slightly better overclocking much less coil line um and when i say slightly better overclocking like we're talking like 10 megahertz um so not really like some like from the factory i guess the biggest like if you were a regular consumer i guess you'd be very excited to have less coil wine from my perspective it's kind of like not really not really that much of a difference um but um yeah so so that that that is why i've been searching for um more appropriate capacitors for doing that with and unfortunately high voltage capacitors um and when i say high voltage i mean like 16 volts because it's an atx power supply that puts 12 volts everything like if it's not like what i consider everything in like everything that goes into a vrm is a high voltage instead of you know output of the vrm low voltage input of the vrm high voltage and technically i i do know there's boost converters but anyway so for input filtering which is you know a 12 volt like input power for a gpu is 12 volts um you need to use 16 volt capacitors because you can't really get like like if they made 14s which they don't you could use 14s but you can't like yeah so 16 volt capacitors problem with 16 volt capacitors is that relative to their you know capacitance and esr they are very expensive like 217 microfarads 16 volts 8 milliohms of esr and this is 5000 hours at 105 degrees rated and one of them is a pound like and this is honestly not that bad this is the cheap option no the expensive option is this and this isn't even the really expensive option this is just sort of the expensive option i'd be living like willing to tolerate um because these guys are 50 million csr these are tantalum polymers yeah see aluminum polymers tantalum polymers um so these have 50 milliamps of esr 16 volts thousand hours at 125 degrees so probably equivalent to 5000 hours at 105 or similar to that i'm not sure how that works out exactly um and the thing is these are one and a half pounds but they only have a hundred microfarads of capacitance so in order to reach like so if i just need like because the 330s i used i added six of them so that was like just under 2000 microfarads extra so if i need two thousand microfarads it's gonna cost me like 22 quid in freaking tantalums or in these it's going to cost me like well a lot less because these are 270 microfarads so i only need like 10 of them and actually 10 of them would already be kind of like way more than than you know like 10 of these is more capacitance than 20 of these this that would be like 27 of these but yeah the thing is i've not yet made up my mind because like obviously these are way more cost effective and also i think like the lower esr is better and like i just generally because my theory because the coil wine happens at the frequency of the frame rate so i really think that the like the the tantalums have a like high frequency advance like high frequency impedance advantage which you can't see it from the esr and it mostly kicks in once you start approaching the megahertz range rather than 100 kilohertz um but the thing is i don't really think the problem here is filtering out things that happen at 100 kilohertz because you're rendering like 200 frames a second which is like 200 hertz right so i don't really like i'm not too concerned about the super high frequency performance and you know through-hole polymers like this are still good to the 100 kilohertz range just fine or 200 kilohertz something like that um so i'm not too worried about so i'm not really like so the whole like oh these are amazing at super high frequencies is like this is not really what i'm looking for here the problem so you might be like wait why are you even considering the tantalums well the thing is these are 6.3 millimeters in diameter so you know going like going back to the hole oh my i square c header fits under the stock back plate none of these don't these very much don't under there's 4.3 millimeters by my measurements there's like 4.3 millimeters of clearance under the back plate um so i can't have 6.3 millimeter caps they're two millimeters too big whereas the tantalums these are only two millimeters in height so these would totally fit but they're also really expensive and probably kind of trash um in terms of well not necessarily trash but i think they probably deliver worse performance um and so yeah i'm really annoyed about like the layout of the pcb because or more like arguably the the heatsink design as well though to be fair the heatsink works really well like i was pushing a ton of power into the card and the heatsink dealt with it like a champ but the thing is like the the design of the heatsink on this card really doesn't leave any space for mods if we zoom out we're going to take it apart now and i've replaced the one of the thermal pads ended up getting absolutely shredded when i was cleaning all the well you know all of these caps i've i used while um when adding all of these caps obviously i was using flux and when i wanted to clean all the flux off well i needed to get all the thermal pads off of the card and you know away so that i could clean the flux off with alcohol which would probably destroy the thermal pads and so in the process of removing the vrm thermal pad i absolutely just destroyed it um it's a very like the power color uses some really nice thermal pads like they're really soft and they seem to have they probably have a really high thermal conductivity the downside to high-end thermal pads is that they're also just like approaching the consistency of thermal paste so they just disintegrate um if you mishandle them um and so those are all gone so like one of the vrms just doesn't have thermal pads anymore it's on thermal putty now um so yeah that's that's what thermal putty looks like it's the white stuff um so that's gonna take a while to like re up well yeah reapply i guess works fine though also cryonaut lh cryonaut extreme edition for the thermal paste on the core works really well um though i'd say you really like this is more like this paste is like i i it's really expensive i think if you're using this on like water cooling or something you're spending unnecessary like spending money for no you're spending way more on thermal pace than you really need to in my opinion um for extreme overclocking however it's basically a must and der bauer sent me a ton of it um so yeah i just kind of ran it on air cooling because it's like the best thermal paste i have so you know may as well um when when pushing the card to its very limits but the problem with this heatsink is that there's no space right like because it's like okay so i want to add like like if i wanted to add extra input filtering capacitors i want them over here and i want them over here right like in this area that's all heatsink there's there's zero space for extra capacitors in that area right so that's sort of my main complaint and the same same goes for the well this and nope the same same issue like we have this part right here which i guess i could like run pretty long well yeah no there's just no space right like that that's the whole like there's really no space for for putting caps on this side of the card um which i'm kind of annoyed about because there's other cards in the past where i've managed to sneak like a lot of capacitors under the heatsink and then this heatsink is just like it is like power color basically went like okay so we're gonna measure the pcb and then we're gonna occupy all of the empty space above it with finstack and it's like well it is a really good heatsink it it can deal with so much heat um i really wish there was more space for capacitors under it though anyway um now so here's the card right from the front um and of course you'll probably oh you can't really tell from from that this angle but actually you can't really tell whatsoever but anyway even on the front so obviously i added a bunch of multi-layer ceramics to the back i also added a bunch to the front i can't help myself i see a like i'm bored i add caps to a card um oh yeah and then the this is the uh extra bulk capacitance for the memory vrm this one doesn't normally exist this one like basically every 6900 xt is missing a capacitor from its memory filter so i put that back and then i put these two over as extra and you can see these are attached with wires because um well the thing is like there's no point desoldering the existing capacitors as far as i'm concerned they're smd so they're perfectly good like they're not panasonic's they're apac but i don't think that's going to make too much of a difference in terms of performance like i don't really like a lot of people do like capacitor replacement uh on motherboards and stuff and i'm more in the the camp of capacitor edition okay because more caps in parallel is better than less caps in parallel um you get less esr less esl and more capacitance like what could possibly go wrong until it tries to turn on and trips short circuit protection which you need to go really far in order to get to that point um but anyway so yeah so i just sort of added those two and then had to use wires because there wasn't really any other way to attach those um but honestly like the this in my experience this like as a sub-optimal as doing this really sh on paper this should be really bad in practice i've not really noticed it causing many issues other than that i've added multi-layer ceramics to the vdd well i added a multi-layer ceramic over here to vddci and then we have v core v core v core more v core over here and more v core over here you might be like why did you leave out these three capacitors and that's because like these memory chips are in the way and i didn't really feel like accidentally resting myself like it would have like in my experience it's not really going like i've not like i've not had a card with dead memory okay like or i've never killed memory chips um memory chips are tough um generally speaking in my experience um but yeah i prefer not to expose them to unnecessary heat so i just kind of avoided those um then from memory we have extra capacitance here and here and those are the 22 microfarads and here and here and there and there unfortunately the card really isn't designed very like like yeah the card really isn't designed to have more capacitance than it comes with from the factory so finding locations to add capacitors was really annoying um and then also more capacitance on vcore over there this is the 0.75 volt rail 1.8 volt rail um vddci memory v core these two phases are soc oh yeah and there's there's a secondary 1.8 volts and i'm not entirely sure that this is an ldo but i think it's an ldo and this is 1.8 volts for the memory um anyway um yeah so input filtering these are the coil wine production inductors these these are as far as i can tell this is where basically all of the coil wine comes from these two guys over here and then the input filtering in terms of bulk capacitance is i'm actually not entirely sure that this is input filtering it might be um it might i hope it is because if it isn't then this is so much worse um but it's basically these four caps and this one and it's like and also these are 180 microfarads each so this has for all intents and purposes no input filtering capacitance and what little capacitance there is is not even really laid out that well because it's like this is the thing you're filtering this is where you put the filter for it really now admittedly these are you know polymer capacitors so they're not meant to deal with high frequency stuff and putting them this far away on the pcb probably doesn't really make that much of a difference but like why are there's why are these why are there why are there so few of them and why are they so small and like because the thing is is like if you compare this to a vega a vega will have a 150 microfarad tantalum polymer capacitor per phase this has six on the entire bloody card and i'm not entirely sure this one's actually for like the vrm over here it might be part of the 5 volt circuit over there like the the 5 volt regulator over here so it's like now yeah this is this is just hugely disappointing to me because if this if this used the same input filtering like design style as what you see on a vega gpu you would have a tantalum polymer capacitor for every single phase there'd be a tent like you'd have one here you'd have them like every single phase would get its own bulk capacitor and then there's this thing where it's like you've got these four and this and that like i i like and it evidently like isn't optimal right like um because it like the card does clock better when you add more input filtering so it's just like yeah so this this definitely well to be fair at stock it's fine but like this is an eighteen hundred quid custom p well custom pcb uh rx 6900 xt it really is just a stretched reference card um there's not really like the thing is there's not that much you could really do to an amd gpu to make it massively better compared to like like yeah you take the like what i would do if i was doing this is like i'd take the reference pcb i'd stretch it like power color did and then i'd add more input filtering capacitors maybe i'd even add them over here i'm not entirely i'm pretty sure you don't actually have to have them directly coming off of directly off of the vrm theoretically it should be a bit better um if you want a very nice example of like distributing your bulk input filtering capacitors look at like asus and gigabyte motherboards well new gigabyte motherboards old gigabyte motherboards didn't believe in that but new gigabyte motherboards are a great example of that uh asus motherboards for many many years like you basically had not necessarily that many capacitors but you did like the capacitors that there were were basically evenly distributed among the phases um around the 12 volt power plane and then there's there's this thing where it's like so this is where our power delivery there's our filter it's like yeah um i have no idea why this doesn't work that way real mystery rx like you know well i didn't prepare an image for this but i i think like so i'm i'm gonna like i wanna drive the point home because i'm annoyed with this i am extremely annoyed i paid 1800 quid for this thing and they couldn't be bothered to put 20 worth of capacitors on it um and also like when i say 20 actually if they put twenty dollars where the the the funny thing is you can't look at the you can't compare the digi q prices that i'm seeing to what like power color would be paying if power color actually decided to design a proper output like proper input filter for this card um so first thing because the thing is i'm being like i'm paying uh you know like not bulk right because i'm buying like 1 or 20 capacitors so i'm not paying bulk prices um and i'm also buying them from digikey i'm not buying them directly from the manufacturer and digi-key and there's distributors that are much cheaper than digikey so you wouldn't even have to go directly to the manufacturer to get a price you just need to get go to a distributor that sells them in like you know a thousand like 10 000 units at a time or something like that just much larger volume um than what like you'd be buying from digikey normally so like you know if if power color put like if i put twenty dollars worth of capacitors on here it's like less than what twenty dollars of capacitors would do if they were uh being applied during the actual manufacturing process of the card um so yeah i'm pretty annoyed especially since like this is the first time i run into a card we're adding input filtering capacitance actually makes a difference okay because normally it doesn't and i guess that makes sense when historically amd gpus would look like this see this look like we're going to get the full size image so we can zoom in on them like there's our input filtering inductors right over here right so wait um okay well there's your input filtering capacitors those are 150 microfarads each and there is one for every single phase even the hbm power phase has its own input filtering capacitor which i think is really excessive because like the hbm doesn't really like at least i don't think it goes through as like violent transient load cycles as like the rest of the dam card does but like this is an input filter what navi comes with is a bloody joke by comparison like let's let's like i'm i'm gonna pull up a card which has like an equivalent uh like a similar input filter to what you get on an avi card okay now admittedly this has a lot less ceramic on it but um one two three four five bulk capacitors for input filtering so yeah that's that's kind of approaching like the six that you get on navi and this is an r7 260 x it's got a single six pin power connector and i think actually these capacitors are like these capacitors are higher capacitance these are 270 microfarads each so this has more input filtering capacitance than a navi than a big navi card like amd what the hell happened um all right let's pull up a reference 7870 because i think this one this is a five phase v core and oops nope oh wait this is the 7850 i clicked on the wrong image there we go so this is a 7870 so you've got four bulk capacitors over there for input filtering you've got two more over here um are they they should be the marker oh yeah so 16 volts i was looking at the wrong end of the capacitor for the voltage rating now so these are 16 volts as well um there's our input filtering inductor um you know and that's just like right around the v core vrm and this is a freaking 7870 another card that comes with two well two six pins okay i was like the single it's not it's two six pins but not much of a difference and then you've got like on the bottom edge i think for the memory yeah you've got two more 16 volt capacitors for the one phase memory power and then another 16 volt capacitor for what i was assuming is the like vddci for the card actually i'm not sure which one of those is which but anyway you get the point okay like how the hell do mid-range cards from 10 years ago have better input filtering than a modern flagship that pulls like twice as much power okay like that's that's the thing is like when i was benching this thing like at stock like the stock actually yeah even the stock power limit on this is like twice the power of a 7 because i believe the 7870 had a 150 watt power limit or something like that and so this pulls twice the freaking power and it has less input filtering like what like how does that make any sense vega had a perfectly good input filter um the 7970 290x fiji like i like this this is like even like even nvidia cards you can't complain about the input filtering configuration on them um not even like historically like it used to be like historic like old nvidia cards had really garbage power delivery but the capacitor configuration was still fine and then we have this thing where it's like okay so we've spent you know we have 70 amp smart power stages we've got these super high end voltage controllers from international rectifier um one of them can do like does 14 phase v core without doublers and two phase soc at the same time um and then you just don't put any input filtering capacitors and it was like but like it's not like they ran out of space right like like i had enough space to put my signature on the stupid thing um so there's evidently plenty of space just like like filtering who needs that like gpus don't have power transients gpus don't pull different amounts of power you know in different parts of the frame render cycle especially not modern ones with their crazy advanced boost algorithms so that you can really maximize opportunistically peak power consumption you know like because that's the thing with really aggressive boost algorithm what's your what your goal is to basically burn as much power when you really need it and then back off of it for the rest of the the workload when as much as possible um and so like so modern gpus are actually even worse in terms of like transient loading on the input side compared to old gpus and it's like yeah um we don't need input filtering capacitors absolutely not not necessary whatsoever and i've just realized this video is an hour long but you know what i think that's a good length um that's that's and that's a that's a good point to stop so yeah um this card so basically i have mixed feelings about the card um vrm is solid you know like the yeah the vrm is solid the output filtering the the thing is is like when i measured this card on an oscilloscope basically every single voltage rail looked really really noisy i've not yet measured it in its most recent uh you know configuration um oh yeah in the some of the like when i was first measuring it it was very noisy and i'm starting to wonder if like maybe it was really noisy because like the voltage every time the card went through a transient the input voltage would dip so hard that it was just like well the like the output can't do anything about the regular like the voltage dipping if the if there's no energy to pull in from the in like from the input side right because so so yeah it's like you know like and it's just like i can't believe this is a thing on a 1800 quid card that pulls 300 like 300 watts that's made by amd like like i don't even get why the reference design is so bad in in the input filtering department like i i i don't get it like maybe the the design is like maybe the assumption was that hey we're gonna just run the card at like the 300 watt power limit and if it never goes above the 300 watt power limit well it's like it's going to be so power limited that the peak power draw is never actually going to get that high right like it'll just always be power throttling to be almost constantly at 300 watts but i'm pretty sure there's some oscilloscope measurements of the power draw of these cards and they don't do that so again it's just like where did the input filter go why does this look like you use like why does this look like a a a 7 8 like not even a seven eight seventy right like but really like not even a seven eight seventy um so yeah um there i'm i'm honestly pretty annoyed like there are some 6900 xt's that fix my complaint about the input filtering but there's a lot of cards that also just take the reference design from amd and stretch it and this this is one really good example of that and i'm i am not impressed um like vrm efficiency is good power handling like the amount of current the vr like in sustained loads this vrm configuration is actually completely fine but the thing is like there's transients this does not work for transients there's not enough capacitance here so um yeah anyway um that's it for the video i guess um thanks for watching and also huge thanks to all of the the patrons and like supporters of the channel because yeah without you guys i wouldn't be able to buy one of these in order to measure that like measure and test that it's actually way more like like that it's honestly kind of disappointing you know like it's not bad just disappointing my my standards might be too high but i feel like when i'm comparing your pcb against a vega i shouldn't have to be complaining about it right like it's like this doesn't even compare to a bloody vega in terms of like this has better power efficiency it's not better filtering and that's like pretty disappointing and i guess somebody might be wondering why don't you just pull the tantalums off of the dead vegas you have well the thing about pulling tantalums off of dead vegas is that i'm pretty sure i burn the out of those caps every time i try to do that and i don't really want to be attaching burnt tantal like if you oh like if you damage a tantalum capacitor they have a tendency to explode so i am not attaching those to a to a card if i don't trust them 110 and i don't trust them 110 in fact i trust them like 50 so yeah they're not going on my card uh not on this one they're going on other stuff maybe but definitely not on this thing so yeah anyway there that is now actually it for the video so thank you for watching like share subscribe leave any comments questions suggestions down in the comment section below and if you'd like to support what i do here with actually hardcore overclocking i have a patreon there's a link to that down in the description below if you check that out it would be much appreciated because uh this thing's expensive yo and then there's also the hoc teespring store where you can pick up shirt stickers posters you know the usual youtuber merch there's a link to that down in the description below as well so it would be much appreciated if you would uh check that out and yeah that's it for the video thanks for watching and goodbye

Channel: Actually Hardcore Overclocking

Views: 36,034

Keywords: Overclocking, PCbuilding, Buildzoid, AHOC, Actually, Hardcore, Hardware, OC

Id: 14f5w4gPcTw

Length: 63min 54sec (3834 seconds)

Published: Mon May 31 2021